Knowledge Discovery from Magnetic Domain Structure Image Using Feature Extended Landau Free Energy Model

Microscopic image data is key information to developing next-generation low-power, high-speed electronic devices. The increasing complexity of nanoscale magnetic domain structures has necessitated advanced analytical techniques to optimize the design of such electronic devices. Traditional visual analysis methods are not only labor-intensive but also highly subjective, often leading to inconsistent and qualitative interpretations. To address these challenges, we have developed an "Extended Landau Free Energy Model" that integrates topology, machine learning, and free energy principles to automate the interpretation of magnetic domain structure data [1-6]. This model enhances our understanding of the mechanisms underlying magnetic interactions and proposes more efficient device structures, contributing significantly to the field of materials science.<br/><br/><b><u>Experiment</u></b><br/>Our approach combines topology and data science with the Landau Free Energy Model to analyze real materials with complex nanostructures. The traditional Landau Free Energy Model, which explains magnetization reversal based on magnetization and magnetic fields, is limited in its applicability to real materials. To overcome this limitation, we utilized persistent homology, a concept from topology, to extract features from complex magnetic domain structures. Interpretable machine learning techniques were then employed to map these features onto a new energy landscape within the information space, resulting in the creation of the "Extended Landau Free Energy Model."<br/>This model allows for the analysis of magnetization reversal processes by correlating changes in magnetic domain structures with energy barriers. Through simple variable transformations and differentiation, the model establishes a bidirectional connection across the hierarchy between micro-scale magnetic domain structures and macro-scale magnetization reversal phenomena. This quantitative analysis enables the identification of physical interactions that govern these processes.<br/><br/><b><u>Results&Discussion</u></b><br/>Applying our model to the analysis of information recording processes in nanomagnetic bodies revealed the dominant role of the demagnetizing field effect. The model successfully visualized the spatial concentration of energy barriers that impede changes in magnetic domain structures, transforming previously undecipherable microscopic data into valuable insights. This result suggests that microscopic data, once considered unusable, can now serve as a rich source of information for device optimization.<br/>Moreover, the model facilitated the proposal of a nanostructure that consumes less energy, highlighting its potential for device design. The model's versatility extends to various materials with complex mechanisms, making it applicable to a wide range of manufacturing fields, including electric vehicle motors and autonomous distributed systems.<br/>The methodology of knowledge discovery along with rule mining, in our model could pave the way for broader impacts in materials research. By automating the interpretation of complex magnetic domain structures, our approach could democratize AI in materials science, enabling more efficient and effective research and development processes.<br/><br/><u><b>References</b></u><br/>[1] A. L. Foggiatto, R. Nagaoka, M. Taniwaki, T. Yamazaki, T. Ogasawara, I. Obayashi, Y. Hiraoka, C. Mitsumata, M. Kotsugi, IEEE Transaction on Magnetics, (2024) accepted<br/>[2] R. Nagaoka, K. Masuzawa, M. Taniwaki, A. L. Foggiatto, T. Yamazaki, I. Obayashi, Y. Hiraoka, C. Mitsumata, M. Kotsugi*, IEEE Transaction on Magnetics, (2024) accepted<br/>[3] S. Kunii, K. Masuzawa, A. Fogiatto, C. Mitsumata, and M. Kotsugi*, Scientific Reports, 12, (2022) 29892<br/>[4] A. Foggiatto, S. Kunii, C. Mitsumata, and M. Kotsugi, Communications Physics, 5 (2022) 277.<br/>[5] S. Kunii, A. Foggiatto, C. Mitsumata, and M. Kotsugi*, Science and Technology of Advanced Materials: Methods, 2 (2022) 445<br/>[6] C. Mitsumata, M. Kotsugi, J. Magn. Soc. Jpn, 46, (2022) pp. 90-96